Approximate Optimum Curbside Utilisation for Pick-Up and Drop-Off (PUDO) and Parking Demands Using Reinforcement Learning

With the uptake of automated transport, especially Pick-Up and Drop-Off (PUDO) operations of Shared Autonomous Vehicles (SAVs), the valet parking of passenger vehicles and delivery vans are envisaged to saturate our future streets. These emerging behaviours would join conventional on-street parking activities in an intensive competition for scarce curb resources. Existing curbside management approaches principally focus on those long-term parking demands, neglecting those short-term PUDO or docking events. Feasible solutions that coordinate diverse parking requests given limited curb space are still absent. We propose a Reinforcement Learning (RL) method to dynamically dispatch parking areas to accommodate a hybrid stream of parking behaviours. A partially-learning Deep Deterministic Policy Gradient (DDPG) algorithm is trained to approximate optimum dispatching strategies. Modelling results reveal satisfying convergence guarantees and robust learning patterns. Namely, the proposed model successfully discriminates parking demands of distinctive sorts and priorities PUDOs and docking requests. Results also identify that when the demand-supply ratio situates at 2:1 to 4:1, the service rate approximates an optimal (83%), and curbside occupancy surges to 80%. This work provides a novel intelligent dispatching model for diverse and fine-grained parking demands. Furthermore, it sheds light on deploying distinctive administrative strategies to the curbside in different contexts.